def resample_feed(self):
"""
random z_noise
"""
z = ops.get_random("normal", self.z_shape)
self.feed.update({self.z_noise: z})
if self.FLAGS.cgan:
if self.c_shape[1] == 34:
c = ops.get_random("getchu_boolean", self.c_shape)
else:
c = ops.get_random("boolean", self.c_shape)
self.feed.update({self.c_noise: c})
def make_fixed_feed(self):
self.fixed_feed = {
self.gen_model.keep_prob: 1.0,
self.disc_model.keep_prob: 1.0,
self.gen_model.training: False,
self.disc_model.training: False
}
self.fixed_noise = ops.get_random("normal", self.z_shape)
if self.FLAGS.cgan:
if self.c_shape[1] == 34:
self.fixed_cond = ops.get_random("getchu_boolean",
self.c_shape)
else:
self.fixed_cond = ops.get_random("boolean", self.c_shape)
self.fixed_feed.update({
self.z_noise: self.fixed_noise,
self.c_noise: self.fixed_cond
})
else:
self.fixed_feed.update({self.z_noise: self.fixed_noise})
def get_noise_batch(self, raw_z, lr, grad):
print("=> noise: min=%f, max=%f" % (raw_z.min(), raw_z.max()))
grad_std = np.std(grad)
noise_list = [raw_z - lr * grad]
for i in range(3):
noise_list.append(raw_z - (2**(i + 1)) * lr * grad)
for i in range(4):
noise_list.append(raw_z - (2**i) * lr * grad +
lr * ops.get_random("normal", raw_z.shape) *
grad_std * 0.1)
noise_list = np.concatenate(noise_list, axis=0)
noise_list = np.maximum(noise_list, -self.RANGE)
noise_list = np.minimum(noise_list, self.RANGE)
return noise_list
def get_z_noise(self):
# origin raw z noise: guassian with 3 sigma
self.origin_raw_z = ops.get_random("normal", (1, 128)) * 3
def get_c_noise(self):
if self.using_cgan == False: return
# raw c noise is gaussian
self.origin_raw_c = ops.get_random("normal", (1, self.n_attrs))